Rebuild t, t_minus_1, t_minus_n and successful from per-context replay metadata through Qwen3VLAgent,
as the RL trainer builds its inputs (Qwen system_prompt_1231, text-action history window 10, one current
screen; UI-Genie window of four prior before-screens plus the current screen). Top-level slices unchanged.
- Goal repair: rows whose goal came from another episode of the same task now use the source trajectory's
goal (error labels, screens and history were all computed on that trajectory); original_goal kept.
- Dropped 172 t/t_minus_1 contexts whose first-error action the replay does not represent (answer 148,
wait/parse failure 22, copy/paste 2), plus goal-mismatch rows whose step count also differs.
- 2,284 rows per slice: t_minus_n restricted to t contexts (all 1,589 kept), successful keeps all 1,802
trajectories; seed 20260924, ids in selection.json; is_next_terminate=True rows never dropped.
- Replay coordinates of GUI-Owl/MA3 steps rebuilt in pixel mode (the manifest's…
is_next_terminate: evidence rule for *Verify tasks (whole history)
*Verify tasks (402 rows across t, t_minus_1, t_minus_n, successful) are relabelled by an evidence rule:
True when the current screen or any earlier history screenshot shows the requested state and no later action
changed that setting; False when no screenshot shows it or the latest one shows the opposite. The judge sees the
whole history (all screenshots, not only the last five), two agreeing passes. Where an earlier state is True and
the next False, the joint cross-check confirmed a state-changing action in between in every case.
Verify ground truth (29 rows): 17/17 True correct, 17/25 successes caught (was 15/15, 15/25).
True now: t 49, t_minus_1 69, t_minus_n 17, successful 77. Annotation files gain verify_* fields;
docs, READMEs, manifests and code (verify_labeler.py, verify_consistency.py, apply_verify.py) updated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UNYnrmb1ku3cno9AxXNYL6
is_next_terminate for successful; goal-mismatch rows always False
successful: 74/2,623 True. 40 are fixed by ground truth (the source policy terminated at the episode's last
step and the AndroidWorld checker accepted it; the three-stage labeler alone caught 33 of those 40), 3 follow from
them through navigation-only steps, the rest passed pass 1, pass 2, the strict gate and cross-checks between
consecutive prompts of the same trajectory.
All slices: rows whose goal differs from their source trajectory (goal_matches_source_trajectory=False) are now
always False. Before this rule the labeler produced vacuous Trues for goals naming items absent from that episode
(10 labels: 9 in successful, 1 in t). t now has 45 True; t_minus_1 and t_minus_n samples are unchanged.
Annotation files gain is_next_terminate_gt_propagated / forced_false_goal_mismatch. Docs, READMEs, manifests and
code (postprocess.py) updated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UNYnrmb1ku3cno9AxXNYL6
is_next_terminate for t, t_minus_1, t_minus_n (successful follows)
New columns in samples.parquet / samples.json metadata: is_next_terminate (the current state already
satisfies the goal, so the next action should be terminate(status="success"); True is high-precision,
False means not confirmed), goal_matches_source_trajectory, is_verify_task. Per-stage decisions and
evidence in <slice>/is_next_terminate_annotations.jsonl. Existing columns and files unchanged.
Labelled with gemini-3.8-flash (two agreeing passes, strict gate, cross-checks between neighbouring
states). Against 276 AndroidWorld outcomes in t: 9/10 True were successes, 234/235 failures labelled
False, 9/41 successes caught. Method, validation and the goal/source-trajectory mismatch (~15% of rows)
in terminate_annotation.md; code in androidflux_tools/next_terminate/.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UNYnrmb1ku3cno9AxXNYL6
30 days ago
Repair all four AndroidFlux contexts: history screens, readable actions and source references
